Apply limit + order optimization in the optimizer (cmu-db#1385)

* Apply limit + order optimization in the optimizer
mtunique · Jun 26, 2018 · 996d240 · 996d240
1 parent f98029f
commit 996d240
Show file tree

Hide file tree

Showing 9 changed files with 151 additions and 42 deletions.
diff --git a/src/include/optimizer/operators.h b/src/include/optimizer/operators.h
@@ -299,9 +299,18 @@ class LogicalDistinct : public OperatorNode<LogicalDistinct> {
 //===--------------------------------------------------------------------===//
 class LogicalLimit : public OperatorNode<LogicalLimit> {
  public:
-  static Operator make(int64_t offset, int64_t limit);
+  static Operator make(
+      int64_t offset, int64_t limit,
+      std::vector<expression::AbstractExpression *> &&sort_exprs,
+      std::vector<bool> &&sort_ascending);
   int64_t offset;
   int64_t limit;
+  // When we get a query like "SELECT * FROM tab ORDER BY a LIMIT 5"
+  // We'll let the limit operator keep the order by clause's content as an
+  // internal order, then the limit operator will generate sort plan with 
+  // limit as a optimization.
+  std::vector<expression::AbstractExpression *> sort_exprs;
+  std::vector<bool> sort_ascending;
 };
 
 //===--------------------------------------------------------------------===//
@@ -470,9 +479,18 @@ class PhysicalOrderBy : public OperatorNode<PhysicalOrderBy> {
 //===--------------------------------------------------------------------===//
 class PhysicalLimit : public OperatorNode<PhysicalLimit> {
  public:
-  static Operator make(int64_t offset, int64_t limit);
+  static Operator make(
+      int64_t offset, int64_t limit,
+      std::vector<expression::AbstractExpression *> sort_columns,
+      std::vector<bool> sort_ascending);
   int64_t offset;
   int64_t limit;
+  // When we get a query like "SELECT * FROM tab ORDER BY a LIMIT 5"
+  // We'll let the limit operator keep the order by clause's content as an
+  // internal order, then the limit operator will generate sort plan with 
+  // limit as a optimization.
+  std::vector<expression::AbstractExpression *> sort_exprs;
+  std::vector<bool> sort_acsending;
 };
 
 //===--------------------------------------------------------------------===//

diff --git a/src/include/planner/order_by_plan.h b/src/include/planner/order_by_plan.h
@@ -28,6 +28,11 @@ class OrderByPlan : public AbstractPlan {
               const std::vector<bool> &descend_flags,
               const std::vector<oid_t> &output_column_ids);
 
+  OrderByPlan(const std::vector<oid_t> &sort_keys,
+              const std::vector<bool> &descend_flags,
+              const std::vector<oid_t> &output_column_ids, const uint64_t limit,
+              const uint64_t offset);
+
   void PerformBinding(BindingContext &binding_context) override;
 
   //===--------------------------------------------------------------------===//
@@ -51,13 +56,21 @@ class OrderByPlan : public AbstractPlan {
     return output_ais_;
   }
 
-  inline PlanNodeType GetPlanNodeType() const override { return PlanNodeType::ORDERBY; }
+  inline PlanNodeType GetPlanNodeType() const override {
+    return PlanNodeType::ORDERBY;
+  }
 
   void GetOutputColumns(std::vector<oid_t> &columns) const override {
     columns = GetOutputColumnIds();
   }
 
-  const std::string GetInfo() const override { return "OrderBy"; }
+  const std::string GetInfo() const override {
+    return std::string("OrderBy") +
+           (limit_
+                ? "(Limit : " + std::to_string(limit_number_) + ", Offset : " +
+                      std::to_string(limit_offset_) + ")"
+                : "");
+  }
 
   void SetUnderlyingOrder(bool same_order) { underling_ordered_ = same_order; }
 

diff --git a/src/optimizer/child_property_deriver.cpp b/src/optimizer/child_property_deriver.cpp
@@ -136,11 +136,17 @@ void ChildPropertyDeriver::Visit(const PhysicalAggregate *) {
                 vector<shared_ptr<PropertySet>>{make_shared<PropertySet>()}));
 }
 
-void ChildPropertyDeriver::Visit(const PhysicalLimit *) {
-  // Let child fulfil all the required properties
-  vector<shared_ptr<PropertySet>> child_input_properties{requirements_};
+void ChildPropertyDeriver::Visit(const PhysicalLimit *op) {
+  // Limit fulfill the internal sort property
+  vector<shared_ptr<PropertySet>> child_input_properties{
+      std::make_shared<PropertySet>()};
+  std::shared_ptr<PropertySet> provided_prop(new PropertySet);
+  if (!op->sort_exprs.empty()) {
+    provided_prop->AddProperty(
+        std::make_shared<PropertySort>(op->sort_exprs, op->sort_acsending));
+  }
 
-  output_.push_back(make_pair(requirements_, move(child_input_properties)));
+  output_.push_back(make_pair(provided_prop, move(child_input_properties)));
 }
 
 void ChildPropertyDeriver::Visit(const PhysicalDistinct *) {

diff --git a/src/optimizer/input_column_deriver.cpp b/src/optimizer/input_column_deriver.cpp
@@ -84,18 +84,40 @@ void InputColumnDeriver::Visit(const QueryDerivedScan *op) {
           output_cols, {input_cols}};
 }
 
-void InputColumnDeriver::Visit(const PhysicalLimit *) { Passdown(); }
+void InputColumnDeriver::Visit(const PhysicalLimit *op) {
+  // All aggregate expressions and TVEs in the required columns and internal 
+  // sort columns are needed by the child node
+  ExprSet input_cols_set;
+  for (auto expr : required_cols_) {
+    if (expression::ExpressionUtil::IsAggregateExpression(expr)) {
+      input_cols_set.insert(expr);
+    } else {
+      expression::ExpressionUtil::GetTupleValueExprs(input_cols_set, expr);
+    }
+  }
+  for (const auto& sort_column : op->sort_exprs) {
+    input_cols_set.insert(sort_column);
+  }
+  vector<AbstractExpression *> cols;
+  for (const auto &expr : input_cols_set) {
+    cols.push_back(expr);
+  }
+  output_input_cols_ =
+      pair<vector<AbstractExpression *>, vector<vector<AbstractExpression *>>>{
+          cols, {cols}};
+}
 
 void InputColumnDeriver::Visit(const PhysicalOrderBy *) {
   // we need to pass down both required columns and sort columns
   auto prop = properties_->GetPropertyOfType(PropertyType::SORT);
   PELOTON_ASSERT(prop.get() != nullptr);
   ExprSet input_cols_set;
   for (auto expr : required_cols_) {
-    if (expression::ExpressionUtil::IsAggregateExpression(expr))
+    if (expression::ExpressionUtil::IsAggregateExpression(expr)) {
       input_cols_set.insert(expr);
-    else
+    } else {
       expression::ExpressionUtil::GetTupleValueExprs(input_cols_set, expr);
+    }
   }
   auto sort_prop = prop->As<PropertySort>();
   size_t sort_col_size = sort_prop->GetSortColumnSize();

diff --git a/src/optimizer/operators.cpp b/src/optimizer/operators.cpp
@@ -443,10 +443,15 @@ Operator LogicalDistinct::make() {
 //===--------------------------------------------------------------------===//
 // Limit
 //===--------------------------------------------------------------------===//
-Operator LogicalLimit::make(int64_t offset, int64_t limit) {
+Operator LogicalLimit::make(
+    int64_t offset, int64_t limit,
+    std::vector<expression::AbstractExpression *> &&sort_exprs,
+    std::vector<bool> &&sort_ascending) {
   LogicalLimit *limit_op = new LogicalLimit;
   limit_op->offset = offset;
   limit_op->limit = limit;
+  limit_op->sort_exprs = std::move(sort_exprs);
+  limit_op->sort_ascending = std::move(sort_ascending);
   return Operator(limit_op);
 }
 
@@ -656,10 +661,15 @@ Operator PhysicalOrderBy::make() {
 //===--------------------------------------------------------------------===//
 // PhysicalLimit
 //===--------------------------------------------------------------------===//
-Operator PhysicalLimit::make(int64_t offset, int64_t limit) {
+Operator PhysicalLimit::make(
+    int64_t offset, int64_t limit,
+    std::vector<expression::AbstractExpression *> sort_exprs,
+    std::vector<bool> sort_ascending) {
   PhysicalLimit *limit_op = new PhysicalLimit;
   limit_op->offset = offset;
   limit_op->limit = limit;
+  limit_op->sort_exprs = sort_exprs;
+  limit_op->sort_acsending = sort_ascending;
   return Operator(limit_op);
 }
 

diff --git a/src/optimizer/plan_generator.cpp b/src/optimizer/plan_generator.cpp
@@ -156,10 +156,35 @@ void PlanGenerator::Visit(const QueryDerivedScan *) {
 }
 
 void PlanGenerator::Visit(const PhysicalLimit *op) {
+  // Generate order by + limit plan when there's internal sort order
+  output_plan_ = std::move(children_plans_[0]);
+  if (!op->sort_exprs.empty()) {
+    vector<oid_t> column_ids;
+    PELOTON_ASSERT(children_expr_map_.size() == 1);
+    auto &child_cols_map = children_expr_map_[0];
+    for (size_t i = 0; i < output_cols_.size(); ++i) {
+      column_ids.push_back(child_cols_map[output_cols_[i]]);
+    }
+
+    PELOTON_ASSERT(op->sort_exprs.size() == op->sort_acsending.size());
+    auto sort_columns_size = op->sort_exprs.size();
+    vector<oid_t> sort_col_ids;
+    vector<bool> sort_flags;
+    for (size_t i = 0; i < sort_columns_size; ++i) {
+      sort_col_ids.push_back(child_cols_map[op->sort_exprs[i]]);
+      // planner use desc flag
+      sort_flags.push_back(!op->sort_acsending[i]);
+    }
+    unique_ptr<planner::AbstractPlan> order_by_plan(new planner::OrderByPlan(
+        sort_col_ids, sort_flags, column_ids, op->limit, op->offset));
+    order_by_plan->AddChild(std::move(output_plan_));
+    output_plan_ = std::move(order_by_plan);
+  }
+
   unique_ptr<planner::AbstractPlan> limit_plan(
       new planner::LimitPlan(op->limit, op->offset));
-  limit_plan->AddChild(move(children_plans_[0]));
-  output_plan_ = move(limit_plan);
+  limit_plan->AddChild(move(output_plan_));
+  output_plan_ = std::move(limit_plan);
 }
 
 void PlanGenerator::Visit(const PhysicalOrderBy *) {
@@ -508,8 +533,8 @@ void PlanGenerator::BuildProjectionPlan() {
 
 void PlanGenerator::BuildAggregatePlan(
     AggregateType aggr_type,
-    const std::vector<std::shared_ptr<expression::AbstractExpression>> *
-        groupby_cols,
+    const std::vector<std::shared_ptr<expression::AbstractExpression>>
+        *groupby_cols,
     std::unique_ptr<expression::AbstractExpression> having_predicate) {
   vector<planner::AggregatePlan::AggTerm> aggr_terms;
   vector<catalog::Column> output_schema_columns;

diff --git a/src/optimizer/query_to_operator_transformer.cpp b/src/optimizer/query_to_operator_transformer.cpp
@@ -111,8 +111,18 @@ void QueryToOperatorTransformer::Visit(parser::SelectStatement *op) {
   }
 
   if (op->limit != nullptr) {
+    const auto &order_info = op->order;
+    std::vector<expression::AbstractExpression *> sort_exprs;
+    std::vector<bool> sort_ascending;
+    for (auto &expr : order_info->exprs) {
+      sort_exprs.push_back(expr.get());
+    }
+    for (auto &type : order_info->types) {
+      sort_ascending.push_back(type == parser::kOrderAsc);
+    }
     auto limit_expr = std::make_shared<OperatorExpression>(
-        LogicalLimit::make(op->limit->offset, op->limit->limit));
+        LogicalLimit::make(op->limit->offset, op->limit->limit,
+                           std::move(sort_exprs), std::move(sort_ascending)));
     limit_expr->PushChild(output_expr_);
     output_expr_ = limit_expr;
   }

diff --git a/src/optimizer/rule_impls.cpp b/src/optimizer/rule_impls.cpp
@@ -811,7 +811,8 @@ void ImplementLimit::Transform(
   const LogicalLimit *limit_op = input->Op().As<LogicalLimit>();
 
   auto result_plan = std::make_shared<OperatorExpression>(
-      PhysicalLimit::make(limit_op->offset, limit_op->limit));
+      PhysicalLimit::make(limit_op->offset, limit_op->limit,
+                          limit_op->sort_exprs, limit_op->sort_ascending));
   std::vector<std::shared_ptr<OperatorExpression>> children = input->Children();
   PELOTON_ASSERT(children.size() == 1);
 

diff --git a/src/planner/order_by_plan.cpp b/src/planner/order_by_plan.cpp
@@ -9,7 +9,7 @@
 // Copyright (c) 2015-17, Carnegie Mellon University Database Group
 //
 //===----------------------------------------------------------------------===//
- 
+
 #include <memory>
 #include <vector>
 
@@ -21,24 +21,35 @@ namespace peloton {
 namespace planner {
 
 OrderByPlan::OrderByPlan(const std::vector<oid_t> &sort_keys,
-		 const std::vector<bool> &descend_flags,
-		 const std::vector<oid_t> &output_column_ids)
- : sort_keys_(sort_keys),
-   descend_flags_(descend_flags),
-   output_column_ids_(output_column_ids) {}
+                         const std::vector<bool> &descend_flags,
+                         const std::vector<oid_t> &output_column_ids)
+    : sort_keys_(sort_keys),
+      descend_flags_(descend_flags),
+      output_column_ids_(output_column_ids) {}
+
+OrderByPlan::OrderByPlan(const std::vector<oid_t> &sort_keys,
+                         const std::vector<bool> &descend_flags,
+                         const std::vector<oid_t> &output_column_ids,
+                         const uint64_t limit, const uint64_t offset)
+    : sort_keys_(sort_keys),
+      descend_flags_(descend_flags),
+      output_column_ids_(output_column_ids),
+      limit_(true),
+      limit_number_(limit),
+      limit_offset_(offset) {}
 
 void OrderByPlan::PerformBinding(BindingContext &binding_context) {
   // Let the child do its binding first
   AbstractPlan::PerformBinding(binding_context);
 
   for (const oid_t col_id : GetOutputColumnIds()) {
-    auto* ai = binding_context.Find(col_id);
+    auto *ai = binding_context.Find(col_id);
     PELOTON_ASSERT(ai != nullptr);
     output_ais_.push_back(ai);
   }
 
   for (const oid_t sort_key_col_id : GetSortKeys()) {
-    auto* ai = binding_context.Find(sort_key_col_id);
+    auto *ai = binding_context.Find(sort_key_col_id);
     PELOTON_ASSERT(ai != nullptr);
     sort_key_ais_.push_back(ai);
   }
@@ -64,39 +75,32 @@ hash_t OrderByPlan::Hash() const {
 }
 
 bool OrderByPlan::operator==(const AbstractPlan &rhs) const {
-  if (GetPlanNodeType() != rhs.GetPlanNodeType())
-    return false;
+  if (GetPlanNodeType() != rhs.GetPlanNodeType()) return false;
 
   auto &other = static_cast<const planner::OrderByPlan &>(rhs);
 
-  // Sort Keys 
+  // Sort Keys
   size_t sort_keys_count = GetSortKeys().size();
-  if (sort_keys_count != other.GetSortKeys().size())
-    return false;
+  if (sort_keys_count != other.GetSortKeys().size()) return false;
 
   for (size_t i = 0; i < sort_keys_count; i++) {
-    if (GetSortKeys()[i] != other.GetSortKeys()[i])
-      return false;
+    if (GetSortKeys()[i] != other.GetSortKeys()[i]) return false;
   }
 
   // Descend Flags
   size_t descend_flags_count = GetDescendFlags().size();
-  if (descend_flags_count != other.GetDescendFlags().size())
-    return false;
+  if (descend_flags_count != other.GetDescendFlags().size()) return false;
 
   for (size_t i = 0; i < descend_flags_count; i++) {
-    if (GetDescendFlags()[i] != other.GetDescendFlags()[i])
-      return false;
+    if (GetDescendFlags()[i] != other.GetDescendFlags()[i]) return false;
   }
 
   // Output Column Ids
   size_t column_id_count = GetOutputColumnIds().size();
-  if (column_id_count != other.GetOutputColumnIds().size())
-    return false;
+  if (column_id_count != other.GetOutputColumnIds().size()) return false;
 
   for (size_t i = 0; i < column_id_count; i++) {
-    if (GetOutputColumnIds()[i] != other.GetOutputColumnIds()[i])
-      return false;
+    if (GetOutputColumnIds()[i] != other.GetOutputColumnIds()[i]) return false;
   }
 
   return AbstractPlan::operator==(rhs);